A pitch extraction algorithm tuned for automatic speech recognition

By Pegah Ghahremani, Bagher BabaAli, Daniel Povey, K. Riedhammer, Jan "Yenda" Trmal, Sanjeev Khudanpur. 2014

350 citations recorded by the source corpus.

Subfields: Artificial Intelligence, Signal Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Signal Processing

Citations in the corpus

UniCATS: A Unified Context-Aware Text-to-Speech Framework with Contextual VQ-Diffusion and Vocoding (2023, background): “This module, akin to the variance adaptor in FastSpeech 2, facilitates conditioning the generation on three-dimensional auxiliary features: pitch, energy, and probability of voice (POV) (Ghahremani et al. 2014).”
VQTTS: High-Fidelity Text-to-Speech Synthesis with Self-Supervised VQ Acoustic Feature (2022, methodology): “The 3dimensional prosody feature is extracted by Kaldi [33].”
A Comparative Study on Transformer vs RNN in Speech Applications (2019, methodology): “The sourceX in ASR is represented as a sequence of 83-dim log-mel filterbank frames with pitch features [20].”
Semantic Mask for Transformer based End-to-End Speech Recognition (2019, methodology): “We represent input signals as a sequence of 80-dim log-Mel filter bank with 3-dim pitch features [18].”
AISHELL-1: An open-source Mandarin speech corpus and a speech recognition baseline (2017, methodology): “icing (POV) feature obtained from Normalized Cross Correlation Function (NCCF), log pitch with POV-weighted mean subtraction over 1.5 second windows, and delta pitch feature computed on raw log pitch [6]. Mean normalization and double deltas are applied on the above features before feeding into the ”
Confidence Estimation and Deletion Prediction Using Bidirectional Recurrent Neural Networks (2018, methodology): “Input features for acoustic models consisted of 24 dimensional ﬁlter bank coefﬁcients, probability of voicing and pitch [32].”
Semi-supervised maximum mutual information training of deep neural network acoustic models (2015, methodology): “For the Babel experiments, we use PLP as the base features, but we additionally append pitch features [36].”
Parsing Speech: a Neural Approach to Integrating Lexical and Acoustic-Prosodic Information (2017, methodology): “word-level vector. The contour features are extracted with 25-ms frames with 10-ms hops using Kaldi (Povey et al., 2011), motivated by the success of the associated f0 features in speech recognition (Ghahremani et al., 2014). Three f0 features are used: warped Normalized Cross Correlation Function (”
Contaminated speech training methods for robust DNN-HMM distant speech recognition (2017, methodology): “The pitch and PoV are estimated through the normalized autocorrelation method discussed in [33].”
ABX-Discriminability Measures and Applications. (Mesures de discriminabilité ABX et applications) (2016, neutral): “Input features consisted of 13 MFCC coefficients plus 3 pitch-related features [222] and their delta and delta-deltas coefficients.”
Semi-Supervised DNN Training with Word Selection for ASR (2017, methodology): “The GMM-HMM features are obtained by splicing +/- 4 frames of the 13-dimensional PLPs (includes C0) extended by 3 kaldi-pitch features [18].”
Advances in Joint CTC-Attention Based End-to-End Speech Recognition with a Deep CNN Encoder and RNN-LM (2017, methodology): “As input features, we used 80 mel-scale filterbank coefficients with pitch features as suggested in [22, 23] for the BLSTM encoder, and adding their delta and delta delta features for the CNN BLSTM encoder [15].”
The Placing Task: A Large-Scale Geo-Estimation Challenge for Social-Media Videos and Images (2014, methodology): “We chose Kaldi pitch [7] because it is readily available and does well with clean audio and SAcC [11] because it does well with noisy audio.”
Semi-Supervised and Cross-Lingual Knowledge Transfer Learnings for DNN Hybrid Acoustic Models Under Low-Resource Conditions (2016, methodology): “We use PLP+pitch features [23] as GMM-HMM front-end.”
Language Independent End-to-End Architecture For Joint Language and Speech Recognition (methodology): “Voxforge often uses the same prompts for several utterances, and we have to avoid the same prompts appeared in the training and test data. implemented in Kaldi [33] for both 8/16 kHz speech signals2.”
Prosody Transfer in Neural Text to Speech Using Global Pitch and Loudness Features (2019, methodology): “We refer readers to [18] for details about the pitch tracking algorithm.”